利用自然语言处理评估青光眼护理的随访模式

摘要(中文译文)

开发一种自然语言处理（NLP）算法，从临床青光眼临床记录的自由文本中识别随访建议，并确定超过推荐随访间隔的患者比例。一项回顾性队列研究。2016年1月至2022年12月在我们的学术眼科诊所接受治疗的成年青光眼患者。结构化的人口统计和诊断数据与临床记录的自由文本一起提取。开发了一种NLP算法，用于从无注释文本中识别推荐的随访间隔。护理失误被定义为在推荐的随访间隔的125%内没有回访。使用泊松分布和对数链的广义估计方程来识别护理失误的风险因素。结果以95%置信区间（CI）的调整后风险比（aRR）报告。自然语言处理算法的性能和识别出护理失误的患者比例。总共评估了221 378份临床记录，代表了6452名青光眼患者的66 228次就诊。NLP算法在66228名患者就诊中确定了47084名患者的随访间隔（71.1%），与金标准图表审查相比表现良好（F1评分0.905，精度0.948，召回率0.867）。超过先前建议随访间隔125%的就诊比例为20.0%（四分位数间距[IQR]：10.0-33.3）。与建议的随访相比，经过的间隔的中位延长时间为52.9天（IQR:22.0-11.5天）。随访中出现任何事件的患者比例为73.7%（4754/6452）。至少有一次护理失误的风险因素包括黑人（aRR=1.29，95%CI:1.22-1.37，与白人相比）和西班牙裔（aRR=1.53，95%CI=1.26-1.87），老年人的风险较低（aRR=0.91，95%CI:0.86-0.97，年龄71至80岁，与≤60岁相比）。在4754名至少有一次护理失误的患者中，1183名（24.9%）在失误后再也没有回来。自然语言处理可以使用个性化的随访间隔可靠地检测青光眼护理中的失误。识别护理失误可能有助于转诊资源，以降低随访损失的风险。专有或商业披露可在本文末尾的脚注和披露中找到。

Abstract

To develop a natural language processing (NLP) algorithm that identifies follow-up recommendations from the free text of clinical glaucoma clinic notes, and to determine the proportion of patients who lapse beyond recommended follow-up intervals. A retrospective cohort study. Adult glaucoma patients treated from January 2016 through December 2022 at our academic ophthalmology practice. Structured demographic and diagnosis data were extracted alongside the free text of clinical notes. An NLP algorithm was developed to identify recommended follow-up intervals from note-free text. Lapses in care were defined as absence of a return visit within 125% of the recommended follow-up interval. Generalized estimating equations with a Poisson distribution and log link were used to identify risk factors for lapsing in care. Results are reported as adjusted risk ratios (aRRs) with 95% confidence intervals (CIs). Natural language processing algorithm performance and proportion of patients with an identified lapse in care. In total, 221 378 clinical notes were assessed, representing 66 228 visits by 6452 unique glaucoma patients. The NLP algorithm identified a follow-up interval for 47 084 of 66 228 (71.1%) patient visits and demonstrated high performance compared to gold standard chart review (F1 score 0.905, precision 0.948, recall 0.867). The proportion of visits that exceeded 125% of the preceding recommended follow-up interval was 20.0% (interquartile range [IQR]: 10.0-33.3). The median excess duration of the lapsed interval compared to recommended follow-up was 52.9 days (IQR: 22.0-111.5 days). The proportion of patients with any event of lapsing in follow-up was 73.7% (4754/6452). Risk factors for having at least 1 lapse in care included Black race (aRR = 1.29, 95% CI: 1.22-1.37, compared to White race) and Hispanic ethnicity (aRR = 1.53, 95% CI: 1.26-1.87), with lower risk for older age (aRR = 0.91, 95% CI: 0.86-0.97, age 71 to 80 years compared to ≤60 years). Among the 4754 patients with at least 1 lapse in care, 1183 (24.9%) never returned after lapsing. Natural language processing can reliably detect lapses in glaucoma care using individualized follow-up intervals. Identifying lapses in care may facilitate referrals to resources to mitigate risk of loss to follow-up. Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

如何引用